GPT-5: OpenAI's August 2025 Reasoning Model, Benchmarks, API Controls, Pricing and Deprecation

What GPT-5 is

GPT-5 is a model from OpenAI, released on 7 August 2025 and offered in the API as gpt-5, with the snapshot gpt-5-2025-08-07. OpenAI released it in three sizes, gpt-5, gpt-5-mini and gpt-5-nano, to let developers trade off performance, cost and latency. In ChatGPT, GPT-5 is a system of reasoning, non-reasoning and router models; in the API, gpt-5 is the reasoning model behind ChatGPT's top performance, and the non-reasoning ChatGPT model is offered separately as gpt-5-chat-latest. It is a proprietary model with no published weights. On LMArena's text leaderboard it has 1435 Elo (rank 108, high effort).

The ChatGPT system

OpenAI described GPT-5 in ChatGPT as a unified system with a fast model that answers most questions, a deeper reasoning model called GPT-5 thinking for harder problems, and a real-time router that picks between them based on conversation type, complexity, tool needs and explicit intent, such as a user asking it to think hard. The router is trained on signals including when users switch models, preference rates and measured correctness. Once usage limits are reached, a mini version of each model handles the remaining queries. GPT-5 was available to all ChatGPT users, with more usage for Plus subscribers and access to GPT-5 pro, a version with extended reasoning, for Pro subscribers. OpenAI also introduced safe completions, a safety-training approach that teaches the model to give the most helpful answer it can within safety boundaries, sometimes answering only partly or at a high level, and to explain why when it refuses.

Reported results

OpenAI reported 74.9% on SWE-bench Verified, up from 69.1% for o3, omitting 23 of 500 problems that did not run reliably on its infrastructure, and 88% on Aider polyglot. It said that relative to o3 at high reasoning effort, GPT-5 used 22% fewer output tokens and 45% fewer tool calls on SWE-bench Verified. At high effort it scored 94.6% on AIME 2025 and 85.7% on GPQA Diamond without tools, and 24.8% on Humanity's Last Exam without tools. OpenAI also said GPT-5 made about 80% fewer factual errors than o3 on prompts from the LongFact and FactScore benchmarks, and answered correctly 89% of the time on BrowseComp Long Context inputs of 128K to 256K tokens. These are OpenAI's own evaluations.

API controls and context

OpenAI added developer controls with the GPT-5 family. A verbosity parameter, with values low, medium and high, steers how long answers are by default, though explicit instructions in the prompt take precedence. The reasoning_effort parameter gained a minimal value, alongside low, medium (the default) and high, to return answers quickly with little reasoning first. Custom tools let the model call tools with plaintext instead of JSON, and can be constrained by context-free grammars supplied by the developer. The model can also send preamble messages before tool calls. In the API, all GPT-5 models accept up to 272,000 input tokens and emit up to 128,000 reasoning and output tokens, for a total context length of 400K tokens. OpenAI's model page lists a knowledge cutoff of 30 September 2024 and support for streaming, function calling and structured outputs.

Pricing and deprecation

gpt-5 is priced at $1.25 per million input tokens and $10 per million output tokens. At launch it was available on the Responses API and Chat Completions API and was the default model in Codex CLI. OpenAI's model page now lists GPT-5 as a previous model and points new work to GPT-6 Astra. On 11 June 2026 OpenAI announced the deprecation of gpt-5 and gpt-5-2025-08-07, with shutdown on 11 December 2026; its deprecations page names gpt-6.1-sol as the replacement.

Official Resources